diff --git a/python/tvm/relax/frontend/torch/exported_program_translator.py b/python/tvm/relax/frontend/torch/exported_program_translator.py
index a521bfa4b7c7..d4fe40806c93 100644
--- a/python/tvm/relax/frontend/torch/exported_program_translator.py
+++ b/python/tvm/relax/frontend/torch/exported_program_translator.py
@@ -1314,7 +1314,8 @@ def _grid_sampler_2d(self, node: fx.Node) -> relax.Var:
         """Convert torch.nn.functional.grid_sample to relax.op.image.grid_sample."""
         args = self.retrieve_args(node)
         data = args[0]
-        grid = args[1]
+        # PyTorch uses [N, H_out, W_out, 2]; Relax NCHW uses [N, 2, H_out, W_out].
+        grid = self.block_builder.emit(relax.op.permute_dims(args[1], axes=[0, 3, 1, 2]))
         interp_mode = args[2] if len(args) > 2 else 0
         pad_mode = args[3] if len(args) > 3 else 0
         align_corners = args[4] if len(args) > 4 else False
diff --git a/tests/python/relax/test_frontend_from_exported_program.py b/tests/python/relax/test_frontend_from_exported_program.py
index ef266b2be05a..300fc7d5f866 100644
--- a/tests/python/relax/test_frontend_from_exported_program.py
+++ b/tests/python/relax/test_frontend_from_exported_program.py
@@ -8727,28 +8727,46 @@ class expected:
         @R.function
         def main(
             input_1: R.Tensor((1, 3, 4, 4), dtype="float32"),
-            grid: R.Tensor((1, 2, 2, 2), dtype="float32"),
-        ) -> R.Tuple(R.Tensor((1, 3, 2, 2), dtype="float32")):
+            grid: R.Tensor((1, 3, 5, 2), dtype="float32"),
+        ) -> R.Tuple(R.Tensor((1, 3, 3, 5), dtype="float32")):
             with R.dataflow():
-                lv: R.Tensor((1, 3, 2, 2), dtype="float32") = R.image.grid_sample(
+                grid_nchw: R.Tensor((1, 2, 3, 5), dtype="float32") = R.permute_dims(
+                    grid, axes=[0, 3, 1, 2]
+                )
+                lv: R.Tensor((1, 3, 3, 5), dtype="float32") = R.image.grid_sample(
                     input_1,
-                    grid,
+                    grid_nchw,
                     method="bilinear",
                     layout="NCHW",
                     padding_mode="zeros",
                     align_corners=True,
                 )
-                gv: R.Tuple(R.Tensor((1, 3, 2, 2), dtype="float32")) = (lv,)
+                gv: R.Tuple(R.Tensor((1, 3, 3, 5), dtype="float32")) = (lv,)
                 R.output(gv)
             return gv
 
     example_args = (
         torch.randn(1, 3, 4, 4, dtype=torch.float32),
-        torch.randn(1, 2, 2, 2, dtype=torch.float32),
+        torch.randn(1, 3, 5, 2, dtype=torch.float32),
     )
     verify_model(GridSample(), example_args, {}, expected)
 
 
+@pytest.mark.parametrize("height,width", [(2, 2), (3, 5)])
+def test_grid_sample_numerically(height, width):
+    class GridSample(Module):
+        def forward(self, data, grid):
+            return torch.nn.functional.grid_sample(data, grid, align_corners=False)
+
+    # In the 2x2 case, the wrong layout still produces the expected shape.
+    data = torch.arange(16, dtype=torch.float32).reshape(1, 1, 4, 4)
+    y, x = torch.meshgrid(
+        torch.linspace(-0.5, 0.5, height), torch.linspace(-0.5, 0.5, width), indexing="ij"
+    )
+    grid = torch.stack((x, y), dim=-1).unsqueeze(0)
+    verify_model_numerically(GridSample(), (data, grid), rtol=1e-6, atol=1e-6)
+
+
 def test_torchvision_roi_align():
     torchvision = pytest.importorskip("torchvision")